Predicting Electrical Vehicle Charging Patterns at Public Charging Stations

As electrical vehicle (EV) technologies become mature, there is a rapid growth in public charging infrastructures. With public charging stations, an accurate prediction of local charging demand can enable many applications, such as dynamic charging allocations, optimizations of power grid operations, and proactive planning for EV users. However, charging demand prediction is a challenging problem, because it is usually affected by the diverse behaviors exhibited by different user groups. In this paper, we explore a data-driven approach to leverage historical charging records for the prediction of future charging demand. We develop a predictive model that distinguishes between the behaviors of both registered long-term users and unregistered short-term users. Utilizing a real-world dataset of 28053 records over 798 days at multiple locations, we employ several supervised learning algorithms to evaluate the performances of our models. Our evaluation results demonstrate that our model, enhanced with the XGBoost, significantly outperforms alternative solutions. It achieves a reduction in prediction error up to 40.8% at the finest time granularity (15-minute interval).